Big Data Processing and Analytics - Mining Noisy Visual Data and Learning Transferrable Predictive Models

点击翻译按钮获取中文摘要

The term ‘Big Data’ has recently emerged to characterize a wide variety of techniques and problems that involve the capture, management, processing, analysis and use of large quantities of data. The increase in our ability to capture, store and process data has reached a point where the impact of Big Data is now seen on the front page of national newspapers. The applicant has extensive experience across a variety of prototypical Big Data problems, including a previous discovery grant on ‘large scale data mining’. Research will focus on developing broadly applicable techniques for visual data processing and analysis through looking at a variety of problems involving visual data with the potential for high impact. Key areas will consist of medical image analysis, object and activity recognition focusing on applications to video indexing, next generation computer animation, intelligent transportation and robotics.Research will focus upon the following common needs, problems, challenges and research questions identified through the applicant’s first hand experience with previous Big Data research projects, namely: 1) The need for acceleration techniques capable of processing large data sets including pre-processing, feature extraction, data modeling, optimization and analysis techniques. 2) The need for principled theory and implementations of techniques that optimize over relevant measures of performance while also accounting for different pre-processing, model complexity, resource constraints and the amount of labelled vs. unlabeled data. 3) The challenges associated with how to most effectively exploit potentially enormous quantities of unlabeled data, as well as weakly or partially and/or noisily labelled data. 4) The problems associated with data collection and reducing human labelling effort. 5) The open questions of how to more effectively transfer learned models or representations obtained using data in one domain to another domain. To achieve these goals we will perform experiments using standard evaluation data sets as well as further develop, curate and create a number of our own data sets covering the themes of faces, emotions, human activities, scene types, objects and medical imagery. In particular, to obtain large quantities of weakly or noisily labelled data and perform experiments on methods for learning with noisy labels we will further develop a large dataset of video that has been annotated by Descriptive Video Services for the blind. We will also collect our own 3D and 4D object and activity recognition data sets centred on the themes of intelligent transportation, robotics and computer animation.Recent research combining large data sets with highly accelerated optimization of deep neural network learning techniques has yielded impressive results on a wide variety of competitive problems. Here we will explore and compare the ways in which other novel deep architectures and other techniques can benefit from the combination of big data and algorithm acceleration. Accelerated algorithms will then be used to develop techniques for the complete optimization of pipelines including both pre-processing steps and hyper-parameters.We also wish to enhance the transferability of models and representation when applied to test data that has been collected in related but different settings. We hypothesize that complete pipeline optimization may lead to more transferrable methods if validation sets representative of the underlying types of domain transfer are used. For visual recognition, we also hypothesize that techniques explicitly accounting for the 4D nature of our world may yield improved transferability. The project will result in the training of highly qualified personnel in the high demand area of Big Data.